Elamur Research⌕
☰
home / papers / 2026 / MIRAGE: Metadata-guided Image Retrieval and Answer Generation for E-commerce Troubleshooting

MIRAGE: Metadata-guided Image Retrieval and Answer Generation for E-commerce Troubleshooting

Paper · 2026 · Rishav Sahay, Lavanya Sita Tekumalla and Anoop Saladi

Summary

Troubleshooting help for consumer electronics is still mostly text-only. Images from brand support sites and product detail pages are rarely attached to knowledge-base (KB) instructions. When they are attached by embedding similarity or by proximity on a page, the image may show the wrong product variant or the wrong component. MIRAGE addresses this by using structured metadata rather than raw pixels as the common ground between text and images. An LLM labels every KB chunk and image with a shared schema: product attributes (brand, model), troubleshooting context, and visual aspects such as ports or buttons. The framework has three modules:

  • M-Link connects images to chunks under strict matching rules.
  • M-Gen is a RAG generator that reads image metadata and inserts image slots into troubleshooting steps.
  • M-Eval is an LLM-as-a-Judge that scores the outputs in the same metadata space.

On enterprise e-commerce data, this approach clearly beats embedding-based linking in both visual coverage and alignment. The result suggests that interpretable metadata can make multimodal grounding scalable and precise.

Key points

  • Data and models: The evaluation covers about 100K KB chunks and 35K images from 10 product types. Metadata is extracted with claude-3-haiku.

  • Results: Compared with embedding-based baselines, MIRAGE achieves over 40 pp higher coverage of high-quality linked visuals. Its linking quality and response quality also improve by more than 45 pp.

  • M-Link rules: A link requires all of the following:

    • the brand matches;
    • the model is consistent, although model-agnostic chunks may match any model;
    • context similarity is at or above a threshold δ;
    • at least one aspect pair is similar at or above ε.

    Near-duplicate images are then removed with CLIP.

  • M-Gen: The generator works only from image metadata, not raw pixels. It places <img_slot> tokens after steps, and these are replaced with actual images when the response is rendered.

  • M-Eval metrics: Responses are scored on Image Relevance, Attribute Alignment, Aspect Alignment and Image Groundedness. Each score is reported as the percentage of successful step–image pairs.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 1Research subjects and languages the work is about.Information retrieval & search · 473 works

Cite

Cite as Rishav Sahay, Lavanya Sita Tekumalla and Anoop Saladi. 2026. MIRAGE: Metadata-guided Image Retrieval and Answer Generation for E-commerce Troubleshooting. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-industry.56

@inproceedings{sahay2026mirage,
  title = {{MIRAGE: Metadata-guided Image Retrieval and Answer Generation for E-commerce Troubleshooting}},
  author = {Rishav Sahay, Lavanya Sita Tekumalla and Anoop Saladi},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-industry.56},
  url = {https://doi.org/10.18653/v1/2026.eacl-industry.56}
}

References

4 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. [5] Junnan Li, Dongxu Li, Caiming Xiong, and Steven CH Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In European Conference on Computer Vision (ECCV), pages 38–56. Springer. [6] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae…
  2. [20] Qinhan Yu, Zhiyou Xiao, Binghui Li, Zhengren Wang, Chong Chen, and Wentao Zhang. 2025. MRAMG-Bench: A Comprehensive Benchmark for Advancing Multimodal Retrieval-Augmented Multimodal Generation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Informat… find
  3. [21] Honglei Zhang and 1 others. 2020. Autoknow: Self-driving knowledge collection for products of thousands of types. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery And Data Mining, pages 2729–2739. find
  4. [22] Deyao Zhu and et al. 2023. Minigpt-4: Enhancing vision-language understanding with gpt-4-level capabilities. arXiv preprint arXiv:2304.10592.

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sahay-2026-mirage-metadata-guided-image-retrieval/